检索结果-内蒙古大学图书馆

MMInstruct: a high-quality multi-modal instruction tuning dataset with extensive diversity

Science China(information Sciences) 2024年第12期67卷 36-51页

作者： Yangzhou LIU Yue CAO Zhangwei GAO Weiyun WANG Zhe CHEN Wenhai WANG Hao TIAN Lewei LU Xizhou ZHU Tong LU Yu QIAO Jifeng DAI School of Computer Science Nanjing University School of Electronic Information and Electrical Engineering Shanghai Jiao Tong University Shanghai AI Laboratory School of Computer Science Fudan University Department of Information Engineering The Chinese University of Hong Kong SenseTime Research Department of Electronic Engineering Tsinghua University

Despite the effectiveness of vision-language supervised fine-tuning in enhancing the performance of vision large language models(VLLMs), existing visual instruction tuning datasets include the following limitations.(1) Instruction annotation quality: despite existing VLLMs exhibiting strong performance,instructions generated by those advanced VLLMs may still suffer from inaccuracies, such as hallucinations.(2) Instructions and image diversity: the limited range of instruction types and the lack of diversity in image data may impact the model's ability to generate diversified and closer to real-world scenarios outputs. To address these challenges, we construct a high-quality, diverse visual instruction tuning dataset MMInstruct,which consists of 973k instructions from 24 domains. There are four instruction types: judgment, multiplechoice, long visual question answering, and short visual question answering. To construct MMInstruct, we propose an instruction generation data engine that leverages GPT-4V, GPT-3.5, and manual correction. Our instruction generation engine enables semi-automatic, low-cost, and multi-domain instruction generation at 1/6 the cost of manual construction. Through extensive experiment validation and ablation experiments,we demonstrate that MMInstruct could significantly improve the performance of VLLMs, e.g., the model fine-tuning on MMInstruct achieves new state-of-the-art performance on 10 out of 12 benchmarks. The code and data shall be available at https://***/yuecao0119/MMInstruct.

关键词： instruction tuning multi-modal multi-domain dataset vision large language model

来源：评论

学校读者我要写书评

暂无评论

Privacy-preserving recommendation with coarse-grained spatiotemporal contexts

引用

Science China(information Sciences) 2025年第4期68卷 66-81页

作者： Lei CHEN Chen GAO Jiahuan LEI Xiaoyi DU Xinlei SHI Hengliang LUO Depeng JIN Yong LI Meng WANG Department of Electronic Engineering BNRist Tsinghua University Meituan Inc. School of Computer Science and Information Engineering Hefei University of Technology

The behavior of users on online life service platforms like Meituan and Yelp often occurs within specific finegrained spatiotemporal contexts(i.e., when and where). Recommender systems, designed to serve millions of users, typically operate in a fully server-based manner, requiring on-device users to upload their behavioral data, including fine-grained spatiotemporal contexts, to the server, which has sparked public concern regarding privacy. Consequently, user devices only upload coarse-grained spatiotemporal contexts for user privacy protection. However, previous research mostly focuses on modeling fine-grained spatiotemporal contexts using knowledge graph convolutional models, which are not applicable to coarse-grained spatiotemporal contexts in privacy-constrained recommender systems. In this paper, we investigate privacy-preserving recommendation by leveraging coarse-grained spatiotemporal contexts. We propose the coarse-grained spatiotemporal knowledge graph for privacy-preserving recommendation(CSKG), which explicitly models spatiotemporal co-occurrences using common-sense knowledge from coarse-grained contexts. Specifically, we begin by constructing a spatiotemporal knowledge graph tailored to coarse-grained spatiotemporal contexts. Then we employ a learnable metagraph network that integrates common-sense information to filter and extract co-occurrences. CSKG evaluates the impact of coarsegrained spatiotemporal contexts on user behavior through the use of a knowledge graph convolutional network. Finally, we introduce joint learning to effectively learn representations. By conducting experiments on two real large-scale datasets,we achieve an average improvement of about 11.0% on two ranking metrics. The results clearly demonstrate that CSKG outperforms state-of-the-art baselines.

关键词： privacy-preserveing coarse-grained spatiotemporal contexts recommender systems

来源：评论

学校读者我要写书评

暂无评论

Enhanced Acceleration for Generalized Nonconvex Low-Rank Matrix Learning

引用

Chinese Journal of electronics 2025年第1期34卷 98-113页

作者： Hengmin Zhang Jian Yang Wenli Du Bob Zhang Zhiyuan Zha Bihan Wen School of Electrical and Electronic Engineering Nanyang Technological University School of Computer Science and Engineering Nanjing University of Science and Technology School of Information Science and Engineering East China University of Science and Technology Department of Electrical and Computer Engineering University of Macau

Matrix minimization techniques that employ the nuclear norm have gained recognition for their applicability in tasks like image inpainting, clustering, classification, and reconstruction. However, they come with inherent biases and computational burdens, especially when used to relax the rank function, making them less effective and efficient in real-world scenarios. To address these challenges, our research focuses on generalized nonconvex rank regularization problems in robust matrix completion, low-rank representation, and robust matrix regression. We introduce innovative approaches for effective and efficient low-rank matrix learning, grounded in generalized nonconvex rank relaxations inspired by various substitutes for the ?0-norm relaxed functions. These relaxations allow us to more accurately capture low-rank structures. Our optimization strategy employs a nonconvex and multi-variable alternating direction method of multipliers, backed by rigorous theoretical analysis for complexity and *** algorithm iteratively updates blocks of variables, ensuring efficient convergence. Additionally, we incorporate the randomized singular value decomposition technique and/or other acceleration strategies to enhance the computational efficiency of our approach, particularly for large-scale constrained minimization problems. In conclusion, our experimental results across a variety of image vision-related application tasks unequivocally demonstrate the superiority of our proposed methodologies in terms of both efficacy and efficiency when compared to most other related learning methods.

关键词： Learning systems Image recognition Minimization Computational efficiency Complexity theory Matrix decomposition Optimization Image reconstruction Singular value decomposition Convergence

来源：评论

学校读者我要写书评

暂无评论

The Guesswork of Ordered Statistics Decoding: Guesswork Complexity and Decoder Design

引用

IEEE Transactions on information Theory 2025年第6期71卷 4167-4192页

作者： Yue, Chentao She, Changyang Vucetic, Branka Li, Yonghui University of Sydney School of Electrical and Computer Engineering NSW Sydney2006 Australia School of Electronic and Information Engineering Shenzhen China

This paper investigates guesswork over ordered statistics and formulates the achievable guesswork complexity of ordered statistics decoding (OSD) in binary additive white Gaussian noise (AWGN) channels. The achievable guesswork complexity is defined as the number of test error patterns (TEPs) processed by OSD immediately upon finding the correct codeword estimate. The paper first develops a new upper bound for guesswork over independent sequences by partitioning them into Hamming shells and applying Hölder’s inequality. This upper bound is then extended to ordered statistics, by constructing the conditionally independent sequences within the ordered statistics sequences. Next, we apply these bounds to characterize the statistical moments of the OSD guesswork complexity. We show that the achievable guesswork complexity of OSD at maximum decoding order can be accurately approximated by the modified Bessel function, which increases exponentially with code dimension. We also identify a guesswork complexity saturation threshold, where increasing the OSD decoding order beyond this threshold improves error performance without further raising the achievable guesswork complexity. Finally, the paper presents insights on applying these findings to enhance the design of OSD decoders. © 1963-2012 IEEE.

关键词： Bessel functions

来源：评论

学校读者我要写书评

暂无评论

A simulation study of polarization characteristics of ultrathin CsPbBr3 nanowires with different cross-section shapes and sizes

引用

Chinese Physics B 2023年第2期32卷 341-346页

作者：杨康胡回清王娇娇邓玲玲陆云清王瑾 School of Telecommunication and Information Engineering Nanjing University of Posts and TelecommunicationsNanjing 210023China School of Opto-Electronic Engineering Nanjing University of Posts and TelecommunicationsNanjing 210023China

The polarization characteristics of ultrathin CsPbBr3nanowires are investigated. Especially, for the height of crosssection of nanowires between 2 nm and 25 nm, the normalized intensity and polarization ratio ρ of CsPbBr3nanowires with triangular, square and hexagonal cross-section shapes are compared. The results show that, along with the increase of the height of cross-section, the polarization ratios of these three nanowires decrease until T = 15 nm, and increase ***, along with the increase of the cross-section area up to 100 nm~2, the polarization ratios of these three nanowires increase too. In general, for the same height or area, the polarization ratio ρ of these nanowires follows ρhexagon> ρsquare>ρtriangle. Therefore, the nanowire with the hexagonal cross-section should be chosen, where for a cross-section height of 2 nm and a length-height ratio of 20 : 1, the maximal polarization ratio is 0.951 at the longitudinal center of the NW. Further,for the hexagonal NW with a cross-section height of 10 nm, the hexagonal NW with a length-height ratio of 45 : 1 exhibits the maximal polarization ratio at the longitudinal center of the NW. These simulation results predict the feasible size and shape of CsPbBr3nanowire devices with high polarization ratios.

关键词： CsPbBr3 nanowire polarization characteristics

来源：评论

学校读者我要写书评

暂无评论

Identify information sources with different start times in complex networks based on sparse observers

引用

Chinese Physics B 2024年第11期33卷 467-479页

作者： Yuan-Zhang Deng Zhao-Long Hu Feilong Lin Chang-Bing Tang Hui Wang Yi-Zhen Huang School of Computer Science and Technology Zhejiang Normal UniversityJinhua 321004China School of Physics and Electronic Information Engineering Zhejiang Normal UniversityJinhua 321004China School of Information Engineering Jinhua PolytechnicJinhua 321016China

The dissemination of information across various locations is an ubiquitous occurrence,however,prevalent methodologies for multi-source identification frequently overlook the fact that sources may initiate dissemination at distinct initial *** there are many research results of multi-source identification,the challenge of locating sources with varying initiation times using a limited subset of observational nodes remains *** this study,we provide the backward spread tree theorem and source centrality theorem,and develop a backward spread centrality algorithm to identify all the information sources that trigger the spread at different start *** proposed algorithm does not require prior knowledge of the number of sources,however,it can estimate both the initial spread moment and the spread *** core concept of this algorithm involves inferring suspected sources through source centrality theorem and locating the source from the suspected sources with linear *** experiments from synthetic and real network simulation corroborate the superiority of our method in terms of both efficacy and ***,we find that our method maintains robustness irrespective of the number of sources and the average degree of *** with classical and state-of-the art source identification methods,our method generally improves the AUROC value by 0.1 to 0.2.

关键词： complex networks information spread source identification backward spread centricity

来源：评论

学校读者我要写书评

暂无评论

An Effective Power Optimization Approach Based on Whale Optimization Algorithm with Two-Populations and Mutation Strategies

引用

Chinese Journal of electronics 2024年第2期33卷 423-435页

作者： Juncai HE Zhenxue HE Jia LIU Yan ZHANG Fan ZHANG Fangfang LIANG Tao WANG Limin XIAO Xiang WANG Hebei Agricultural University Beijing Information Science and Technology University School of Computer Science and Engineering Beihang University School of Electronic and Information Engineering Beihang University

Power is an issue that must be considered in the design of logic circuits. Power optimization is a combinatorial optimization problem, since it is necessary to search for a logical expression that consumes the least amount of power from a large number of Reed-Muller(RM) logical expressions. The existing approach for optimizing the power of multi-output mixed polarity RM(MPRM) logic circuits suffer from poor optimization results. To solve this problem, a whale optimization algorithm with two-populations strategy and mutation strategy(TMWOA) is proposed in this paper. The two-populations strategy speeds up the convergence of the algorithm by exchanging information about the two-populations. The mutation strategy enhances the ability of the algorithm to jump out of the local optimal solutions by using the information of the current optimal solution. Based on the TMWOA, we propose a multi-output MPRM logic circuits power optimization approach(TMMPOA). Experiments based on the benchmark circuits of the Microelectronics Center of North Carolina(MCNC) validate the effectiveness and superiority of the proposed TMMPOA.

关键词： Multi-output mixed polarity Reed-Muller Power optimization Combinatorial optimization problem Whale optimization algorithm Two-populations strategy Mutation strategy

来源：评论

学校读者我要写书评

暂无评论

Autonomous embodied navigation task generation from natural language dialogues

引用

Science China(information Sciences) 2025年第5期68卷 119-132页

作者： Haifeng XU Yongchang LI Lumeng MA Chunwen LI Yanzhi DONG Xiaohu YUAN Huaping LIU Department of Automation Tsinghua University School of Physics and Electronic Information Yantai University School of Electrical and Electronic Engineering Shanghai Institute of Technology Department of Computer Science and Technology Tsinghua University

Robots are increasingly being deployed in densely populated environments, such as homes, hotels, and office buildings, where they rely on explicit instructions from humans to perform tasks. However, complex tasks often require multiple instructions and prolonged monitoring, which can be time-consuming and demanding for users. Despite this, there is limited research on enabling robots to autonomously generate tasks based on real-life scenarios. Advanced intelligence necessitates robots to autonomously observe and analyze their environment and then generate tasks autonomously to fulfill human requirements without explicit commands. To address this gap, we propose the autonomous generation of navigation tasks using natural language dialogues. Specifically, a robot autonomously generates tasks by analyzing dialogues involving multiple persons in a real office environment to facilitate the completion of item transportation between various *** propose the leveraging of a large language model(LLM) through chain-of-thought prompting to generate a navigation sequence for a robot from dialogues. We also construct a benchmark dataset consisting of 625 multiperson dialogues using the generation capability of LLMs. Evaluation results and real-world experiments in an office building demonstrate the effectiveness of the proposed method.

关键词： proactive robot robot navigation service robot large language model

来源：评论

学校读者我要写书评

暂无评论

Data delivery delay and cross-layer packet size analysis for reliable transmission of Licklider transmission protocol in space networks

引用

Science China(information Sciences) 2024年第9期67卷 229-244页

作者： Guannan YANG Ruhai WANG Kanglian ZHAO Wenfeng LI Dong YAN School of Information Engineering Nanjing University of Finance and Economics Drayer Department of Electrical and Computer Engineering Lamar University School of Electronic Science and Engineering Nanjing University Beijing Aircraft Overall Design Department

Delay/disruption tolerant networking(DTN) is proposed as a networking architecture to overcome challenging space communication characteristics for reliable data transmission service in presence of long propagation delays and/or lengthy link disruptions. Bundle protocol(BP) and Licklider Transmission Protocol(LTP) are the main key technologies for DTN. LTP red transmission offers a reliable transmission mechanism for space networks. One of the key metrics used to measure the performance of LTP in space applications is the end-to-end data delivery delay, which is influenced by factors such as the quality of spatial channels and the size of cross-layer packets. In this paper, an end-to-end reliable data delivery delay model of LTP red transmission is proposed using a roulette wheel algorithm, and the roulette wheel algorithm is more in line with the typical random characteristics in space networks. The proposed models are validated through real data transmission experiments on a semi-physical testing platform. Furthermore, the impact of cross-layer packet size on the performance of LTP reliable transmission is analyzed, with a focus on bundle size, block size, and segment size. The analysis and study results presented in this paper offer valuable contributions towards enhancing the reliability of LTP transmission in space communication scenarios.

关键词： Licklider transmission protocol DTN bundle protocol cross-layer packets size space communication

来源：评论

学校读者我要写书评

暂无评论

Event-Driven Attention Network:A Cross-Modal Framework for Efficient Image-Text Retrieval in Mass Gathering Events

引用

computers, Materials & Continua 2025年第5期83卷 3277-3301页

作者： Kamil Yasen Heyan Jin Sijie Yang Li Zhan Xuyang Zhang Ke Qin Ye Li School of Computer Science and Engineering University of Electronic Science and Technology of ChinaChengdu611731China School of Information and Software Engineering University of Electronic Science and Technology ofChinaChengdu611731China Kashi Institute of Electronics and Information Industry Kashi844508China

Research on mass gathering events is critical for ensuring public security and maintaining social ***,most of the existing works focus on crowd behavior analysis areas such as anomaly detection and crowd counting,and there is a relative lack of research on mass gathering *** believe real-time detection and monitoring of mass gathering behaviors are essential formigrating potential security risks and ***,it is imperative to develop a method capable of accurately identifying and localizing mass gatherings before disasters occur,enabling prompt and effective *** address this problem,we propose an innovative Event-Driven Attention Network(EDAN),which achieves image-text matching in the scenario of mass gathering events with good results for the first *** image-text retrieval methods based on global alignment are difficult to capture the local details within complex scenes,limiting retrieval *** local alignment-based methods aremore effective at extracting detailed features,they frequently process raw textual features directly,which often contain ambiguities and redundant information that can diminish retrieval efficiency and degrade model *** overcome these challenges,EDAN introduces an Event-Driven AttentionModule that adaptively focuses attention on image regions or textual words relevant to the event *** calculating the semantic distance between event labels and textual content,this module effectively significantly reduces computational complexity and enhances retrieval *** validate the effectiveness of EDAN,we construct a dedicated multimodal dataset tailored for the analysis of mass gathering events,providing a reliable foundation for subsequent *** conduct comparative experiments with other methods on our dataset,the experimental results demonstrate the effectiveness of *** the image-to-text retrieval task,EDAN achieved the best performance on the R@5 metric,w

关键词： Mass gathering events image-text retrieval attention mechanism

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：